Catastrophic Forgetting
Fine-tuning on a new task quietly erodes capabilities the model already had, because every weight is free to move.
Catastrophic forgetting is what happens when full fine-tuning, free to adjust every weight, optimizes so hard for the new task’s data that it degrades performance on things the base model used to do well — with no signal during training that it’s happening, since nothing in the loss function protects unrelated capabilities. It’s one of the reasons LoRA, which freezes the base weights, tends to preserve general capability better than full fine-tuning.